Greenberg and Witten

by Carson Reynolds

“Adaptive personalized interfaces-A question of viability” seeks to answer affirmatively the question posed in the title. That is, adaptive interfaces are viable. Here viable means narrowly, can perform better than their static counterparts.

The article begins with a definition of adaptive personalized interfaces. A user model is first defined as “a set of rules which a computer system follows to determine its reaction to a user.” The authors strike a distinction between personalization (when a user model is unique for each user or group) and adaptive (when the model is altered during interaction with the user).

The authors then present three types of user modeling:

(1) system designers adjust system to fit current needs in response to feedback from the population.

(2) let user modify the working environment themselves

(3) system monitors the user’s activity and tries to adapt automatically.

The paper concerns itself with the last, which it calls adaptive user modeling.

A literature review occurs in the next section in which the authors summarize various pros and cons for adaptation. Briefly:


Variations in user expertise
Evolving user needs
User has appropriate control
Minimize user-designer conflicts

Dynamics of user-system concurrent modeling
User does not have appropriate control
Difficulty of implementation
Inaccuracies of model construction

Outside of these characteristics the authors note that in the literature they looked at no one had done a comparative evaluation.

The paper moves on to discuss the development of an adaptive phone directory. After collecting some statistical data about phone usage, the authors argue phone dialing patterns follow a Zipf distribution.

The authors do a uniform-subdivision tree directory and compare that with a tree sorted on popularity. Users were then asked to use the two menu systems (the adaptive one, after it was “primed”).

The null hypothesis is that personalization doesn’t affect selections speed or error rate and that the order of the presentation of the two systems has no effect on performance.

After running 26 volunteers through the study an ANOVA was performed. The adaptive design and task ordering were found to be significant with an alpha 0.01. Error rates were also found to be lower.

The paper provides an “existence proof” of adaptive user interfaces that perform well, as opposed to a general design principle. Some domains are ill-suited to the adaptive approach, the authors concede.

There’s a glaring problem with this paper though: namely, users didn’t start from “scratch” with the system. Basically, they proved a tailored system could help users perform the task well, but they didn’t look at the effect of inconsistent interaction on the user. Namely: the user memorizes a particular menu structure and then it changes.